|Publication number||US4651289 A|
|Application number||US 06/460,623|
|Publication date||Mar 17, 1987|
|Filing date||Jan 24, 1983|
|Priority date||Jan 29, 1982|
|Also published as||DE3381386D1, EP0085545A2, EP0085545A3, EP0085545B1|
|Publication number||06460623, 460623, US 4651289 A, US 4651289A, US-A-4651289, US4651289 A, US4651289A|
|Inventors||Kenichi Maeda, Tsuneo Nitta|
|Original Assignee||Tokyo Shibaura Denki Kabushiki Kaisha|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (9), Non-Patent Citations (1), Referenced by (59), Classifications (21), Legal Events (6)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This invention relates to an apparatus and method for recognizing input patterns such as voice patterns and character patterns. Pattern recognition is gaining more acceptance as a fundamental technique for inputting information into computer systems. One method called the similarity method or pattern matching method is a well-known pattern recognition process and is widely utilized in the field of character recognition. Several different similarity methods are known, including the simple similarity method, multiple similarity method and mixed similarity method.
The simple similarity method utilizes a single separate reference pattern, for each category, represented by a vector which is stored in a dictionary memory. A single reference pattern corresponding to each designated category represents, for example, a certain character or voice pattern to be identified. That is, one category can consist of the letter (A), another category can consist of lower case letter (a). In voice recognition, separate categories can consist of the respective sounds for pronouncing each of the vowels (a, e, i, o and u). These reference patterns are then compared with the vector representations of the patterns to be identified (i.e., input patterns) to determine its numerical value of similarity. A high value of similarity indicates that the input pattern is identical or nearly identical to the reference pattern. In particular, the simple similarity method can be performed as follows. First, signals representing the input pattern are sampled and these discrete sampled values are stored as vector components of the input signal. This input vector is then compared with the vectors representing each category. A numerical value of similarity is then calculated for each category which indicates the degree of similarity between the input pattern and the reference pattern for each category. Second, the maximum value of similarity is determined from all the calculated values; this value thus identifies the category to which the input patterns belong.
This simple similarity method has an advantage in that the design of the dictionary of reference patterns can be easily automated, and is not greatly affected by such local noise as stain or scratches in the patterns. It is liable to be affected adversely, however, by such overall changes in the patterns which occur in handwritten letters or voice patterns. That is due to the wide variation in handwriting and voice patterns or pronounications, more deformations in the input pattern can occur. Thus, it is impractical to represent each category by a single reference pattern.
Consequently, other methods have been devised to recognize the input pattern in view of such wide deformations. One such method is the multiple similarity method as disclosed in U.S. Pat. No. 3,688,267 and the mixed similarity method as disclosed in U.S. Pat. No. 3,906,446.
According to the multiple similarity method, a plurality of reference pattern vectors are created for each category. The multiple similarity for a certain category is defined as the sum of the square root of the values of simple similarity between the input pattern and every reference pattern in the same category. As in the case of simple similarity discussed above, recognition is carried out as follows. First, signals representing the input pattern are sampled and these discrete sampled values are stored as vector components of the input signal. This input vector is then compared with each reference pattern vector in the same category. A numerical value of similarity is then calculated for each comparison; the square root of these values are then summed to provide a multiple similarity value for each category. Second, the maximum value of similarity is detected from all calculated values; this value thus identifies the category to which the input pattern belongs.
In the case of mixed similarity, the procedures discussed above for multiple similarity are employed. In addition, the similarity values for mutually similar reference patterns are identified and subtracted to provide even more accurate identification.
The above-described multiple similarity and mixed similarity methods are useful to recognize patterns capable of having numerous variations or overall deformations. However, the conventional systems employing such methods require storage of numerous reference patterns to provide sufficient data to accurately identify and recognize various input patterns. In fact, it is very costly and time consuming to compile the necessary data. Not only is an unduly large memory capacity needed, but excessive computer time is required to calculate the numerous matrix calculations needed to analyze and compare the various stored reference pattern data and input patterns. Consequently, preparation and computation of reference patterns stored in a computer memory for achieving complete and accurate recognition of patterns subject to various deformations have been impractical. As a result, many systems have been developed with a limited reference pattern storage to avoid the cost and incident problems discussed above; consequently, misrecognition has frequently occured when input patterns subject to various deformations have been applied. The industry, therefore, has required a system which can easily be adapted and tailored for special and individual needs without developing an unduly large common memory of reference patterns.
It is therefore an important object of the invention to provide a new and improved pattern recognition apparatus and method.
It is a further object of the invention to provide a new and improved recognition apparatus and method capable of recognizing, with high accuracy, patterns capable of having various deformations.
It is a more specific object of the invention to provide a new and improved pattern recognition apparatus and method with an improved dictionary to thereby provide accurate recognition for patterns which were once misrecognized.
A further object of the present invention is to provide a flexible pattern recognition system which can be easily tailored to many applications without the necessity of constructing a costly and unduly large common memory of reference patterns.
A still further object of the present invention is to accurately identify input patterns capable of various deformations without requiring excessive matrix calculations and computer time.
According to the invention, there is provided a pattern recognition apparatus including a vector generating unit for generating an input vector representing the characteristics of an unknown input pattern, a dictionary unit which stores a plurality of reference vectors for each category, a similarlity calculating unit which calculates a similarity between the input vector and a plurality of reference vectors for each category, a comparing unit which determines the category to which the input pattern belongs by comparing the similarities derived from the similarity calculating unit, and an additional dictionary generating unit which generates additional reference vectors for a particular need or application.
The vector generating unit comprises an input unit for converting the unknown input pattern into electrical signals, and a pre-processing unit receiving the electrical signals and generating the input vector which represents the characteristics of the unknown input pattern.
The dictionary unit has a common dictionary memory for storing common reference vectors previously prepared and an additional dictionary memory for storing additional reference vectors generated by the additional dictionary generating unit.
The similarity calculating unit calculates similarities between the input vector and reference vectors stored in both the common and additional dictionary for each category. The similarity for each category is calculated as follows. First, scalar (inner) products between the input vector and each of the common reference vectors belonging to each category are calculated. The scalar products thus obtained are squared. The squared scalar products are summed to provide a first sum. Second, scalar products between the input vector and each additional reference vector corresponding to each category are calculated. These scalar products are likewise squared. Then these second squared scalar products are summed to provide a second sum. The similarity for each category is obtained by adding the first sum and the second sum for each category.
The additional dictionary generating unit generates an additional reference vector obtained by subtracting the components of at least common reference vectors, in the specified category, from the input vector. The additional reference vector is stored in the storage area within the additional dictionary memory corresponding to a specified category.
Thus, it is possible to satisfy the objective mentioned above. Other objects and features of this invention will be apparent from the following description read in connection with the accompanying drawings.
FIG. 1 is a block diagram of the pattern recognition apparatus showing an embodiment of this invention.
FIG. 2 is a block diagram of the input unit and pre-processing unit of FIG. 1.
FIG. 3A is a block diagram of the similarity calculating unit of FIG. 1.
FIG. 3B is a diagram of the calculator 52-1 of FIG. 3A.
FIG. 3C is a diagram of the calculator 53-1 of FIG. 3A.
FIG. 3D is a diagram of the absolute value squaring circuit 54 of FIG. 3A.
FIG. 4 is a block diagram of the comparing unit of FIG. 1.
FIG. 5A is a block diagram of the additional dictionary generating unit of FIG. 1.
FIG. 5B is a flow chart showing the operation of the additional dictionary generating unit of FIG. 5A.
FIG. 5C is a block diagram of the vector generator 70 of FIG. 5A.
FIG. 6 is another embodiment of a pre-processing unit of FIG. 1.
FIG. 7 is a flow chart showing another operation of the additional dictionary generating unit of FIG. 5A.
FIG. 1 shows the preferred embodiment of the apparatus according to the present invention for recognizing voice patterns. A theoretical explanation will be given with reference to FIG. 1 in order to clarify the features of the invention.
Input unit 1 converts an input voice into electrical signals and is then converted into a pattern vector in a pre-processing unit 2. The input pattern vector (hereinafter indicated by F) represents the characteristics of the input voice signal. The vector F consists of N number of components fi (i=1,2, - - - , N), each component, for example, being a frequency spectrum component of the input voice.
A common dictionary memory 3 stores a plurality of reference vectors for each category which have been previously prepared as reference patterns for permitting multiple or mixed similarity as mentioned above. A category l (l=1,2, - - - L) which is a subject of recognition includes a set of references vectors indicated by [φm.sup.(l) ] or indicated individually by vectors φ1.sup.(l), φ2.sup.(l), - - - , φM (l).sup.(l). The value M.sup.(l) is the number of reference vectors in the category l. Each reference vector consists of N number of components; that is, the same dimension as input vector F. The vectors in this category should preferably satisfy an orthogonal relation given by the following formula: ##EQU1## where (φm.sup.(l), φm'.sup.(l) denotes a scalar (inner) product between vectors φm.sup.(l) and φm'.sup.(l). Based upon this formula, the number of reference vectors required to represent a given category is minimized so that the size of common dictionary memory 3 is reduced.
An additional dictionary memory 4 stores a plurality of additional reference vectors which are generated by an additional dictionary generating unit 7 described later. The set of additional reference vectors for the category l is indicated by ψn.sup.(l), and each additional reference vector is indicated by ψ1.sup.(l), ψ2.sup.(l), - - - , ψN (l).sup.(l), where N.sup.(l) is a number of additional reference vectors for the category l stored in memory 4. Similar to vector F, each additional reference vector also consists of N number of components.
A similarity calculating unit 5 calculates similarities for individual categories. The calculation of a similarity for one category includes three steps, In the first step, the similarity S.sub.[F].sup.(l) between the input vector F and reference vector φm.sup.(l) of category l is calculated as follows: ##EQU2## In this equation, (F, φm.sup.(l)) denotes a scalar product between vectors F and φm.sup.(l), am.sup.(l) (m=1,2, - - - , M.sup.(l)) denotes a coefficient corresponding to the vector φm.sup.(l) (m=1,2, - - - , M.sup.(l)), and |F| is the absolute value of the input vector F defined as follows:
Further, each reference vector φm.sup.(l) is normalized so that |φm.sup.(l) |=1 in equation (1). According to the conventional multiple or mixed similarity method mentioned above, similarities for all categories obtained by equation (1) are compared with each other to determine the category to which the input pattern belongs.
In the second step, the similarity calculating unit 5 of this invention further calculates the similarity S'.sub.[F].sup.(l) between the input vector F and additional reference vectors ψm.sup.(l) which are generated by an additonal dictionary generating unit 7. The similarity S'.sub.[F].sup.(l) is calculated as follows: ##EQU3## In this equation, bn.sup.(l) (n=1,2, - - - , N.sup.(l)) denotes a coefficient corresponding to the vector ψn.sup.(l) (n=1,2, - - - , N.sup.(l)), and each additional reference vector ψn.sup.(l) is normalized so that |ψn.sup.(l) |=1.
In the third step, unit 5 calculates the summed similiarity T.sub.[F] of the input vector F for each category as follows:
T.sub.[F] =S.sub.[F] +S'.sub.[F] (3)
The similarities T.sub.[F].sup.(1), T.sub.[F].sup.(2), - - - , T.sub.[F].sup.(L) for each category are supplied to a comparing unit 6. Comparing unit 6 detects the maximum value among the L summed similarities T.sub.[F].sup.(l) (l=1,2, - - - , L), and determines the category to which the input pattern belongs. Comparing unit 6 produces an output signal representing a name or code of the category. In some cases wherein the maximum similarity is very close to another similarity, comparing unit 6 produces an output reject signal which establishes that the apparatus cannot recognize or identify the input pattern. The output of comparing unit 6 is preferably displayed on a display unit 8 in order to show the result to an operator.
The additional reference vectors ψn.sup.(l) are generated in additional dictionary unit 7 in the following manner. Assume that the input pattern belongs to the category A, and that this input contains a deformation, for example, a deformation, representing a peculiarity in the operator's pronounciation. Upon being supplied with and an input pattern, similarity calculating unit 5 might produce an output indicating that a similarity S.sub.[F].sup.(B) for another category (i.e., incorrect category B) is greater than the similarity S.sub.[F].sup.(A) for the correct category A as shown by the following expression:
In that situation, the additional dictionary unit 7 is manually or automatically energized and generates an additional vector ψ1.sup.(A) in accordance with the following formula: ##EQU4##
Equation (4) shows that the additional vector ψ1.sup.(A) is obtained by calculations using input vector F, previously recognized as an incorrect category, and reference vectors [φm.sup.(A) ]. The vector thus obtained is stored in the additional dictionary memory 4 as one of the additional vectors belonging to the category A (i.e., correct category). In other words, vector ψ1.sup.(A) represents a vector wherein the particular deformation components of the reference vectors of category A are removed from the input vector F. The vector ψ1.sup.(A) given by the equation (4) satisfies the orthogonal relationship with vectors [φm.sup.(A) ], and |ψ1.sup.(A) |=1. After storing additional vector ψ1.sup.(A) into the additional dictionary memory 4, the similarity calculating unit 5 calculates the similarity T.sub.(F).sup.(A) for category A in accordance with equation (3). The similarity T.sub.(F).sup.(A) is now found to be greater than S.sub.(F).sup.(A) to the extent calculated from equation (1) by the amount ##EQU5## Therefore, it is possible to make the similarity T.sub.(F).sup.(A) for category A greater than the similarity T.sub.(F) for the category B by giving coefficient b1.sup.(A) an appropriate positive value. Thus, input patterns including particular deformations can be correctly recognized and proper categories identified by storing the additional vectors into the proper categories of additional dictionary memory 4.
It is also possible, if desired, to store an additional vector ψ1.sup.(B) into category B instead of category A. In this case, the additional dictionary generating unit 7 generates the vector ψ1.sup.(B) as follows: ##EQU6## Vector ψ1.sup.(B) is stored in additional dictionary memory 4 as one of the additional vectors belonging to the category B. Vector ψ1.sup.(B), given by equation (5), also satisfies the orthogonal relationship with vectors [φm.sup.(B) ], and |ψ1.sup.(B) |=1.
After the storage of vector ψ1.sup.(B), the similarity for the category B is calculated by the unit 5 in accordance with equation (3). It differs from the similarity which would result without the calculation and storage of ψ1.sup.(B) by the amount ##EQU7## Therefore, it is possible to make the similarity T.sub.(F).sup.(B) for the category B smaller than the similarity T.sub.(F).sup.(A) for the category A by giving coefficient b1.sup.(A) an appropriate negative value. Further, it may also be desirable to generate two vectors ψ1.sup.(A), ψ1.sup.(B) according to equations (4), (5) and store both vectors in the additional dictionary memory 4.
The apparatus, according to FIG. 1, has two different operation modes, mode I and mode L. Mode I performs an identification or recognition process where uknown input patterns are recognized; mode L performs a learning or storing process where additional reference vectors are generated and stored in the additional dictionary memory so that the system can be tailored to special needs or applications.
FIG. 2 shows the configuration of the input unit 1 and pre-processing unit 2 in FIG. 1. Those elements are similar to the ones disclosed in copending U.S. application Ser. No. 366,667, filed Apr. 8, 1982, now U.S. Pat. No. 4,503,557. A microphone 11 converts the voice signals uttered by the operator into electrical signals. The electrical signals are amplified by the amplifier 12 and supplied to an A/D converter 13. A/D converter 13 converts the electrical signals into digital signals, for example, every 100 usec. The digital signals are supplied to band pass filters (BPF1 -BPF16), each consisting of well known digital filters and extracting the frequency spectrum components in different frequency ranges. The output of each bandpass filter is supplied to squaring circuits 22. Each squaring circuit squares the output of the corresponding bandpass filter in order to obtain the energy component of the input voice.
The output of each squaring circuit 22 is supplied to low pass filters 23; each low pass filter comprises well known digital filters for obtaining the total energy components corresponding to each frequency range. The output of each low pass filter 23 is stored in an input vector register 24 as respective components fn (n=1,2, - - - , 16) of input vector F. Thus, input vector F, representing the characteristic of the input pattern, is extracted as a distribution of energies.
FIG. 3A shows the configuration of the similarity calculating unit 5. Shown is L number of similarity calculating circuits 50-1, 50-2, - - - , 50-l, - - - , 50-L, each corresponding to the category 1, 2, - - - , l, - - - , L. Since each circuit consists of the same configuration, only circuit 50-1 is shown in detail in FIG. 3A.
When the system is in mode I, the input vector F, stored in register 24, is supplied to a buffer register 51. The content of buffer register 51 is supplied to each circuit 50-1, - - - , 50-L. Also supplied to circuit 50-1 are reference vectors [φm.sup.(l) ] and coefficients [am.sup.(l) ] which are stored in memory area 31 of the common dictionary memory 3 and corresponding to category l, and additional reference vectors [ψn.sup.(l) ] and coefficients [bn.sup.(l) ] which are stored in a memory area 41 of the additional dictionary memory 4 and corresponding to category l. Circuit 50-l includes M.sup.(l) number of calculators 52-1, - - - , 52-M.sup.(l), each having the same configuration. Calculator 52-1 is shown in FIG. 3B, as receiving input vector F, reference vector φ1.sup.(l) and coefficient a1.sup.(l). A scalar product circuit 522 calculates the scalar product between input vector F and the reference vector φ1.sup.(l). The output of the scalar product circuit 522 is applied to a squaring circuit 523. The squaring circuit 523 squares and output of scalar product circuit 522. The output of squaring circuit 523 is supplied to a multiplier 524 for multiplying this output by the coefficient a1.sup.(l). The output of multiplier 524 is thus a1.sup.(l).sub.(F,φ1.sup.(l)). Other calculators 52-2, - - - , 52-M.sup.(l) have the same configuration as calculator 51-1.
Circuit 50-1 also includes Nmax number of calculators 53-1, - - - , 53-Nmax, where Nmax is a maximum number of additional reference vectors capable of being stored in memory area 41. Calculator 53-1 is shown in FIG. 3C, as receiving input vector F, reference vector ψ1.sup.(l) and coefficient b1.sup.(l). A scalar product circuit 532 calculates the scalar product between the input vector F and the reference vector ψ1.sup.(l). The output of the scalar product circuit 532 is supplied to a squaring circuit 533. Squaring circuit 533 squares the output of scalar product circuit 532. The output of squaring circuit 533 is supplied to a multiplier 534 for multiplying this output by the coefficient b1.sup.(l) ; the output of the multiplier 534 is thus b1.sup.(l) (F, ψ1.sup.(l)). Other calculators 53-2, - - - , 53-Nmax have the same configuration as calculator 53-1. The outputs of the calculators 52-1, - - -, 52-M.sup.(l) and calculators 53-1, - - - , 53-Nmax are supplied to the adder 55, and their sum is thereby obtained.
The output of adder 55 is ##EQU8## where N.sup.(l) is the number of additional reference vectors stored thus far in the additional memory for the category l. While the memory area 41 of the additional dictionary memory 4 is capable of storing Nmax number of additional reference vectors and coefficients, the residual area of the memory area 41, not being used, stores all zeros.
Circuit 50-1 further includes an absolute value squaring circuit 54 and a divider 56. Absolute value squaring circuit 54 comprises a scalar product circuit 542 as shown in FIG. 3D. The scalar product circuit 542 calculates the scalar product of two inputs; in this case, since the same input vector F is supplied, |F|2 =(F, F). Divider 56 divides the output of the adder 55 by the output of the absolute value squaring circuit 54; as a result, the output of divider 56 is the similarity T.sub.(F) for the category and is obtained as follows: ##EQU9## Thus, circuits 50-1, - - - , 50-L calculate the similarities T.sub.(F).sup.(l), - - - , T.sub.(F).sup.(L), and supply them to the comparing unit 6.
The block diagram of comparing unit 6 is shown in FIG. 4. The similarities T.sub.(F).sup.(l), - - - , T.sub.(F).sup.(L) are supplied to an editing circuit 61. Circuit 61 pairs each calculated similarity with a category name (code) obtained from the category code generator 63, and arranges the pairs in order according to the magnitude of the similarity. Then the ordered pairs are stored in a memory 62. Now, it is assumed that the category code which gives the maximum value T.sub.(F).sup.(A) of the similarities is A and that the category code which give the second largest value T.sub.(F).sup.(B) of the similarities is B.
A decision logic 64 circuit determines the category to which the input pattern belongs by analyzing the contents of memory 62. For example, predetermined threshold values φ1 and φ2 are stored in decision logic circuit 64. Decision logic cicuit 64 receives the maximum value T.sub.(F).sup.(A) and the second largest value T.sub.(F).sup.(B) from memory 62 and compares T.sub.(F).sup.(A) with θ1. In particular, it determines whether T.sub.(F).sup.(A) is larger than the predetermined value θ1, that is:
T.sub.(F).sup.(A) >θ1 (6a)
and also compares whether the difference between T.sub.(F).sup.(A) and T.sub.(F).sup.(B) is larger than the predetermined value θ2, that is:
T.sub.(F).sup.(A) -T.sub.(F).sup.(B) >θ2 (6b)
When the equations (6a), (6b) are satisfied, decision logic circuit 64 determines that the input pattern whose characteristics are represented by the input vector F belong to the category A and outputs the category code A as a recognition result. If one of the equations (6a), (6b) is unsatisfied, decision logic circuit 64 outputs a reject signal indicating that the category cannot be determined.
The outputs of decision logic are supplied to the display unit 8, which comprises a well-known CRT display device and displays the result to the operator. If the operator sees that the input voice pattern is rejected, he can utter the voice pattern again. If the operator finds that the recognition result is incorrect (i.e., the input pattern was misrecognized, or that the particular voice pattern was rejected several times) he will push a mode key (not shown) on keyboard 9. Upon pushing the mode key, the system is set to mode L (Learning mode) in accordance with the following embodiment whereby an additional dictionary generating process is performed.
FIG. 5A shows the configuration of the additional dictionary generating unit 7. As shown, a signal is generated by the keyboard 9 in response to pushing of the mode key. This sets a mode flipflop 71 corresponding to mode L while the rest condition of mode flipflop 71 corresponds to mode I. Unit 7 includes a controller 72 comprising, for example, a micro-computer. Controller 72 operates according to the flow chart shown in FIG. 5B. Controller 72 checks mode flipflop 71 to determine whether it is in a set or reset condition. When the mode flipflop 71 is in the reset condition (mode I), controller 72 repeats the checking process. If controller 72 detects that mode flipflip 71 is in the set condition (mode L), it executes an additional dictionary generating process as outlined in FIGS. 5A and 5B.
During the first step, the operator inputs from keyboard 9 a category code for the additional reference vector he desires to store. Preferably, a message indicating that a category code is needed is sent to the display unit 8 via line 77. Responding to the message command displayed on unit 8, the operator inputs the desired category code. According to the key operation by the operator, keyboard 9 supplies the selected category code (hereinafter designated as A) to controller 72 via line 76.
During the second step, controller 72 by referring to an address table memory 73 fetches the necessary data and supplies it to a vector generator 70 and coefficient calculator 74. Address table memory 73 stores an address table whose data format is shown below: ##STR1## wherein, field l: category code (name) l.
field AD1: start address of the memory area 31 in the common dictionary memory 3 corresponding to the category.
field M.sup.(l) : the number of reference vectors φm.sup.(l) for the category l.
field AD2: start address of the memory area 41 in the additional dictionary memory 4 corresponding to the category l.
field N.sup.(l) : the number of additional reference vectors n.sup.(l) for the category l previously stored.
Controller 72 searches address table memory 73 and obtains each field corresponding to category code A. Controller 72 fetches data within the memory area of the common dictionary memory 3 designated by the field AD1, including reference vectors φm.sup.(l), and coefficients am.sup.(l). It also fetches data within the memory area of the additional dictionary memory designated by the field AD2, including additional reference vectors ψn.sup.(l) and coefficients bm.sup.(l). It also fetches the input vector F from register 24. Then, controller 72 supplies input vector F, reference vectors φm.sup.(l), and additional reference vectors ψn.sup.(l) to vector generator 70, while supplying the coefficents am.sup.(l), bn.sup.(l) to the coefficient calculator 74.
During the third step, vector generator 70 generates a new additional reference vector ψx.sup.(A) and calculator 74 generates a coefficient bx.sup.(A). FIG. 5C shows the configuration of the vector generator 70. Register 701 stores the input vector F supplied by controller 72. Register 702-1, - - - , 702-M' store the reference vectors φm.sup.(A) and additional reference vectors ψn.sup.(A) supplied by controller 72. The number M' of the registers 702-1, - - - , 702-M' is selected as follows:
M'=Mmax +Nmax -1
where Mmax is the maximum value among M.sup.(l) (l=1,2, - - - ,L), and Nmax is the number of additional reference vectors which can be stored in the additional dictionary memory 4. Since the number of reference vectors φm.sup.(A) and additional reference vectors ψn.sup.(A) fetched by controller 72 is M.sup.(A) +N.sup.(A), and M.sup.(A) +N.sup.(A) <M', controller 72 supplies not only the reference vectors and additional reference vectors to the M.sup.(A) +N.sup.(A) number of registers but also "0" to the residual registers. Vector generator 70 also includes scalar product circuits 703-1, - - - , 703-M' (FIG. 5C). Each scalar product circuit calculates a scalar product between the input vector F in register 701 and corresponding reference vectors and additional reference vectors. There is also provided discrete groups of multiplier circuits 704-1, - - - , 704-M', each corresponding to respective scalar product circuits 703-1, - - - , 703-M'. Each multiplier group consists of the number of multipliers equaling the number of components of the vector. For example, multiplier 704-1 consists of 16 multipliers, each multiplying the output of the scalar product circuit 703-1 by a respective component of the reference vector φ1.sup.(A) in register 702-1.
The outputs of the multiplier groups 704-1, - - - , 704-M are supplied to a group of adder circuits group 705, consisting of 16 discrete adder circuits. Each adder of adder group 705 calculates the sum of M' inputs which are the multiplied output of multiplier groups 704 and the same component order of the vectors. For example, the first adder circuit adds the first multiplied component of multiplier 704-1, the first multiplied component of multiplier 704-2 etc. The output of adder group 705 is indicated by a vector D and is given by: ##EQU10## As can be seen from equation (7), vector D has components corresponding to the angles between input vector F and each of the reference vectors. For example, as known from vector mathematics (F,φm.sup.(A))=|F| |φm | Cos θ where θ is the angle between F and φm.sup.(A).
A subtractor group 706 subtracts vector D from the input vector F in register 701. The output of subtractor group 706 is supplied to an absolute value circuit 707 and a divider circuit group 708. The absolute value circuit 707 comprises an absolute value squaring circuit. Thus, circuit can comprise a scalar product circuit, such as the circuit shown in FIG. 3D, and a square root circuit for calculating the square root of the output of the scalar product circuit. The divider circuit group 708 divides the output of the subtractor circuit group 706 by the output of the absolute value circuit 707, and the result is supplied to a register 709. The content of register 709 is the additional vector ψx.sup.(A), which is given by: ##EQU11##
This additional vector ψx.sup.(A) generated by vector generator 70 satisfies an orthogonal relationship not only with reference vectors φm.sup.(A) but also with additional vectors ψn.sup.(A) previously stored in additional dictionary memory 4. Equation (4) shown above, is obtained by substituting N.sup.(A) =0 in equations (7), (8).
The coefficient calculator 74 detects the maximum value Cmax and the minimum value Cmin from the coefficients am.sup.(A), bn.sup.(A) supplied by controller 72, and determines the coefficient bx.sup.(A) as follows: ##EQU12##
During the fourth step, controller 72 receives vector ψx.sup.(A) from vector generator 70 and writes it into the memory area in the additional dictionary memory 4 corresponding to category A as an additional reference vector ψN.spsb.(A)+1.sup.(A). Controller 72 also receives coefficient bx.sup.(B) from the coefficient calculator 74 and writes it into the memory area in the additional dictionary memory 4 corresponding to category A as coefficient bN.spsb.(A)+1.sup.(A). During the last step, controller 72 increments the content of the field N.sup.(A) of address table memory 73 corresponding to category A, and supplies a reset signal to the mode flipflop 71 via line 78, so that the mode condition returns to mode I (see FIG. 5B).
Many scalar product circuits are utilized in this embodiment. It is noted, in general, that the scalar product between a vector P and a vector Q is calculated as follows: ##EQU13## where pi (i=1,2 - - - ,I) are components of the vector P, and q (i=1,2, - - -,I) are components of the vector Q. Therefore, a scalar product circuit can be constructed by utilizing a multiplier and an adder (or accumulator).
FIG. 6 shows another embodiment of the pre-processing unit 2. Provided are band pass filters 25 similar to those shown in FIG. 2; however, only four band pass filters 25 are utilized. As a result, the frequency range is selected to be four times broader than the bandpass filters 21 shown in FIG. 2. The output of each bandpass filter 25 is supplied to a corresponding squaring circuit 26 and a lowpass filter 27. The outputs of lowpass filters 27 are distributed into a register 24 by a distributor 28 at intervals of, for example, 10 msec. The input vector stored in register 24 is indicated by F' whose components are F'(i=1,2, - - - ,16). Components f'1, - - - ,f'4 represent an energy distribution of the first time interval, f'5, - - - ,f'8 represent an energy distribution of the second time interval, and so on. The type of input vector obtained by the current shown in FIG. 2 is effective for the recognition of the voice patterns such as vowel patterns, while the type of input vector obtained by the circuit shown in FIG. 6 is effective for the recognition of consonant patterns or word patterns.
In recognizing character patterns, hand-written or printed on paper, a photo-electric converter such as a CCD scanner can be utilized as the input unit. Such a scanner scans the character patterns and provides electrical signals representing the darkness of each picture elements, so that the components of the input vector corresponds to the darkness of each picture element of a character pattern, for example, as shown in the aforementioned U.S. Pat. No. 3,906,446.
Although each additional vector generated by vector generator 70 is stored in the additional dictionary memory 40 for the above embodiment, this is not always necessary to practice this invention. The additional dictionary generating unit 7 can be modified as discussed below. In particular, a vector memory can be connected to the controller 72 (not shown in FIG. 5A) for storing the generated additional vectors, the coefficients and the number of generated vectors for each category denoted by R.sup.(A) for the category A. The flow chart for controller 72 for operation of this modified system is shown in FIG. 7.
The first, second and third steps shown in FIG. 7 are similar to the steps shown in FIG. 5B. During the fourth step in FIG. 7, however, the additional vector ψx.sup.(A) and the coefficient bx.sup.(A) are stored in the vector memory, and the number R.sup.(A) is incremented by one. During the fifth step, R.sup.(A) is compared with a predetermined value R0. If R.sup.(A) ≠R0, the mode condition is changed to mode I without storing additional vectors. If the condition R.sup.(A) =R0 is satisfied, controller 72 processes as follows.
During the sixth step, vectors ψxr.sup.(A) (r=1,2, - - - , R0) in the vector memory as substituted into the following equation and correlation matrix H is obtained: ##EQU14## where wr.sup.(A) are weighting factors, and <,> denotes the operation of dyad. During the seventh step, eigenvalues are obtained for the correlation matrix as follows:
μn.sup.(A) ψn.sup.(A) =H.sup.(A) ψn.sup.(A) (10)
Eigenvectors are then obtained corresponding to the eigenvalues. In the eighth step, several of the eigenvectors which have the largest eigenvalues are selected and stored in the additional dictionary 4 as additional reference vectors. During the ninth step, N.sup.(A) of the address table 73 is added to the number of selected eigenvectors. The vectors ψxr.sup.(A) are cleared from the vector memory and R.sup.(A) is set to zero. The number of additional vectors to be stored is generally selected so that it is smaller than the number of deformed patterns which were mis-recognized or rejected.
Further, the steps of storing coefficients in the dictionary memory is not necessary when, for example, the following reference vectors are used: ##EQU15##
The above described arrangement is merely illustrative of the principles of the present invention. Numerous modifications and adaptations thereof will be readily apparent to those skilled in the art without departing from the spirit and scope of the present invention.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US3191150 *||Oct 30, 1962||Jun 22, 1965||Ibm||Specimen identification system with adaptive and non-adaptive storage comparators|
|US3509280 *||Nov 1, 1968||Apr 28, 1970||Itt||Adaptive speech pattern recognition system|
|US3688267 *||Nov 2, 1970||Aug 29, 1972||Intern Tarde & Ind Japan||Pattern identification systems operating by the multiple similarity method|
|US3816722 *||Sep 28, 1971||Jun 11, 1974||Nippon Electric Co||Computer for calculating the similarity between patterns and pattern recognition system comprising the similarity computer|
|US3906446 *||Aug 7, 1974||Sep 16, 1975||Taizo Iijima||Pattern identification system|
|US4069393 *||Dec 11, 1974||Jan 17, 1978||Threshold Technology, Inc.||Word recognition apparatus and method|
|US4181821 *||Oct 31, 1978||Jan 1, 1980||Bell Telephone Laboratories, Incorporated||Multiple template speech recognition system|
|US4256924 *||Nov 19, 1979||Mar 17, 1981||Nippon Electric Co., Ltd.||Device for recognizing an input pattern with approximate patterns used for reference patterns on mapping|
|US4319221 *||May 27, 1980||Mar 9, 1982||Nippon Electric Co., Ltd.||Similarity calculator comprising a buffer for a single input pattern feature vector to be pattern matched with reference patterns|
|1||*||A Real Time Spoken Word Recognition System with Various Learning Capabilities of Speaker Differences, System.Computers.Controls, vol. 9, No. 3, Jun. 1978, Nakagawa.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US4805218 *||Apr 3, 1987||Feb 14, 1989||Dragon Systems, Inc.||Method for speech analysis and speech recognition|
|US4821333 *||Aug 22, 1986||Apr 11, 1989||Environmental Research Inst. Of Michigan||Machine learning procedures for generating image domain feature detector structuring elements|
|US4827522 *||Sep 25, 1986||May 2, 1989||Kabushiki Kaisha Toshiba||Pattern recognition apparatus for determining a reference pattern by K-L expansion utilizing a power method|
|US4860360 *||Apr 6, 1987||Aug 22, 1989||Gte Laboratories Incorporated||Method of evaluating speech|
|US4864622 *||Oct 28, 1987||Sep 5, 1989||Sanyo Electric Co., Ltd.||Voice recognizing telephone|
|US4885697 *||Sep 1, 1988||Dec 5, 1989||E. I. Du Pont De Nemours And Company||Method of identifying spectra|
|US4887304 *||Mar 16, 1989||Dec 12, 1989||Raytheon Company||Library image optimization|
|US4901362 *||Aug 8, 1988||Feb 13, 1990||Raytheon Company||Method of recognizing patterns|
|US4908864 *||Apr 2, 1987||Mar 13, 1990||Sharp Kabushiki Kaisha||Voice recognition method and apparatus by updating reference patterns|
|US4926491 *||Jun 6, 1988||May 15, 1990||Kabushiki Kaisha Toshiba||Pattern recognition device|
|US4979213 *||Jul 12, 1989||Dec 18, 1990||Kabushiki Kaisha Toshiba||Orthogonalized dictionary speech recognition apparatus and method thereof|
|US5001760 *||Oct 6, 1988||Mar 19, 1991||Kabushiki Kaisha Toshiba||Speech recognition apparatus and method utilizing an orthogonalized dictionary|
|US5033088 *||Dec 21, 1989||Jul 16, 1991||Voice Processing Corp.||Method and apparatus for effectively receiving voice input to a voice recognition system|
|US5050224 *||May 15, 1990||Sep 17, 1991||Kabushiki Kaisha Toshiba||Character recognition apparatus|
|US5146503 *||Aug 24, 1988||Sep 8, 1992||British Telecommunications Public Limited Company||Speech recognition|
|US5159637 *||Jan 16, 1992||Oct 27, 1992||Fujitsu Limited||Speech word recognizing apparatus using information indicative of the relative significance of speech features|
|US5164992 *||Nov 1, 1990||Nov 17, 1992||Massachusetts Institute Of Technology||Face recognition system|
|US5181256 *||Dec 24, 1990||Jan 19, 1993||Sharp Kabushiki Kaisha||Pattern recognition device using a neural network|
|US5197008 *||Jan 25, 1991||Mar 23, 1993||Mitsubishi Jidosha Kokyo Kabushiki Kaisha||System for controlling the output power of a motor vehicle|
|US5214717 *||Feb 25, 1991||May 25, 1993||Fujitsu Limited||Pattern recognition data processing device using an associative matching method|
|US5278942 *||Dec 5, 1991||Jan 11, 1994||International Business Machines Corporation||Speech coding apparatus having speaker dependent prototypes generated from nonuser reference data|
|US5315668 *||Nov 27, 1991||May 24, 1994||The United States Of America As Represented By The Secretary Of The Air Force||Offline text recognition without intraword character segmentation based on two-dimensional low frequency discrete Fourier transforms|
|US5315689 *||Dec 21, 1992||May 24, 1994||Kabushiki Kaisha Toshiba||Speech recognition system having word-based and phoneme-based recognition means|
|US5317741 *||May 10, 1991||May 31, 1994||Siemens Corporate Research, Inc.||Computer method for identifying a misclassified software object in a cluster of internally similar software objects|
|US5428707 *||Nov 13, 1992||Jun 27, 1995||Dragon Systems, Inc.||Apparatus and methods for training speech recognition systems and their users and otherwise improving speech recognition performance|
|US5428788 *||Jun 4, 1993||Jun 27, 1995||Siemens Corporate Research, Inc.||Feature ratio method for computing software similarity|
|US5438676 *||Jun 4, 1993||Aug 1, 1995||Siemens Corporate Research, Inc.||Method for adapting a similarity function for identifying misclassified software objects|
|US5440742 *||Apr 6, 1994||Aug 8, 1995||Siemens Corporate Research, Inc.||Two-neighborhood method for computing similarity between two groups of objects|
|US5461698 *||Jun 17, 1993||Oct 24, 1995||Siemens Corporate Research, Inc.||Method for modelling similarity function using neural network|
|US5469512 *||Sep 7, 1993||Nov 21, 1995||Sony Corporation||Pattern recognition device|
|US5485621 *||Apr 6, 1994||Jan 16, 1996||Siemens Corporate Research, Inc.||Interactive method of using a group similarity measure for providing a decision on which groups to combine|
|US5590048 *||Jun 4, 1993||Dec 31, 1996||Fujitsu Limited||Block exposure pattern data extracting system and method for charged particle beam exposure|
|US5850627 *||Jun 26, 1997||Dec 15, 1998||Dragon Systems, Inc.||Apparatuses and methods for training and operating speech recognition systems|
|US5909666 *||Jun 26, 1997||Jun 1, 1999||Dragon Systems, Inc.||Speech recognition system which creates acoustic models by concatenating acoustic models of individual words|
|US5915236 *||Jun 26, 1997||Jun 22, 1999||Dragon Systems, Inc.||Word recognition system which alters code executed as a function of available computational resources|
|US5920836 *||Jun 26, 1997||Jul 6, 1999||Dragon Systems, Inc.||Word recognition system using language context at current cursor position to affect recognition probabilities|
|US5920837 *||Jun 26, 1997||Jul 6, 1999||Dragon Systems, Inc.||Word recognition system which stores two models for some words and allows selective deletion of one such model|
|US5983179 *||Jun 26, 1997||Nov 9, 1999||Dragon Systems, Inc.||Speech recognition system which turns its voice response on for confirmation when it has been turned off without confirmation|
|US6038343 *||Feb 5, 1997||Mar 14, 2000||Hewlett-Parkard Company||Character recognition method and apparatus using writer-specific reference vectors generated during character-recognition processing|
|US6073097 *||Jun 26, 1997||Jun 6, 2000||Dragon Systems, Inc.||Speech recognition system which selects one of a plurality of vocabulary models|
|US6092043 *||Jun 26, 1997||Jul 18, 2000||Dragon Systems, Inc.||Apparatuses and method for training and operating speech recognition systems|
|US6101468 *||Jun 26, 1997||Aug 8, 2000||Dragon Systems, Inc.||Apparatuses and methods for training and operating speech recognition systems|
|US6272238 *||Sep 23, 1996||Aug 7, 2001||Canon Kabushiki Kaisha||Character recognizing method and apparatus|
|US6275799 *||Feb 2, 1995||Aug 14, 2001||Nec Corporation||Reference pattern learning system|
|US6285785||Jun 7, 1993||Sep 4, 2001||International Business Machines Corporation||Message recognition employing integrated speech and handwriting information|
|US7035788 *||Apr 25, 2000||Apr 25, 2006||Microsoft Corporation||Language model sharing|
|US7246292 *||Mar 30, 2005||Jul 17, 2007||Fujitsu Limited||Apparatus and method for bit pattern learning and computer product|
|US7895031||Mar 31, 2006||Feb 22, 2011||Microsoft Corporation||Language model sharing|
|US8046363 *||Jan 10, 2007||Oct 25, 2011||Lg Electronics Inc.||System and method for clustering documents|
|US9153230 *||Oct 23, 2012||Oct 6, 2015||Google Inc.||Mobile speech recognition hardware accelerator|
|US20060156103 *||Mar 30, 2005||Jul 13, 2006||Fujitsu Limited||Apparatus and method for bit pattern learning and computer product|
|US20060173674 *||Mar 31, 2006||Aug 3, 2006||Microsoft Corporation||Language model sharing|
|US20070244915 *||Jan 10, 2007||Oct 18, 2007||Lg Electronics Inc.||System and method for clustering documents|
|US20070250532 *||Apr 21, 2006||Oct 25, 2007||Eastman Kodak Company||Method for automatically generating a dynamic digital metadata record from digitized hardcopy media|
|US20150199963 *||Oct 23, 2012||Jul 16, 2015||Google Inc.||Mobile speech recognition hardware accelerator|
|USRE36041 *||Nov 16, 1994||Jan 12, 1999||Massachusetts Institute Of Technology||Face recognition system|
|CN101055585B||Mar 5, 2007||Jan 2, 2013||Lg电子株式会社||System and method for clustering documents|
|EP0357010A2 *||Aug 30, 1989||Mar 7, 1990||E.I. Du Pont De Nemours And Company||Method of identifying spectra|
|EP0357010A3 *||Aug 30, 1989||Sep 18, 1991||E.I. Du Pont De Nemours And Company||Method of identifying spectra|
|U.S. Classification||704/244, 382/190, 704/239|
|International Classification||G06F17/27, G06K9/66, G10L11/00, G10L15/00, G10L15/10, G06K9/62, G06T7/00, G06K9/68|
|Cooperative Classification||G10L15/00, G06F17/2735, G06F17/2715, G06K9/68, G06K9/66|
|European Classification||G10L15/00, G06K9/68, G06F17/27D, G06F17/27A4, G06K9/66|
|Jan 24, 1983||AS||Assignment|
Owner name: TOKYO SHIBAURA DENKI KABUSHIKI KAISHA; 72, HORIKA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST.;ASSIGNORS:MAEDA, KENICHI;NITTA, TSUNEO;REEL/FRAME:004087/0912
Effective date: 19830113
|Sep 7, 1990||FPAY||Fee payment|
Year of fee payment: 4
|Aug 30, 1994||FPAY||Fee payment|
Year of fee payment: 8
|Oct 6, 1998||REMI||Maintenance fee reminder mailed|
|Mar 14, 1999||LAPS||Lapse for failure to pay maintenance fees|
|May 25, 1999||FP||Expired due to failure to pay maintenance fee|
Effective date: 19990317